Apparatus and method for measuring substance concentration based on spectrum learning

Abstract

A method for generating a concentration prediction model for predicting the concentration of a target substance by machine learning is disclosed. The method for generating a concentration prediction model includes a learning data generation step of generating an optimally transformed spectrum as learning data by transforming a basic spectrum of a substance whose concentration is known based on a predetermined transformation condition, and a concentration prediction model generation step of generating a concentration prediction model by machine learning the optimally transformed spectrum transformed based on the predetermined transformation condition generated in the learning data generation step and the actual concentrations of the substances corresponding to the optimally transformed spectrum, thereby suppressing spectral changes due to compounds other than the analyte to be predicted and maximizing spectral changes corresponding to the concentration of the analyte, thereby improving accuracy in predicting the concentration of the substance.

Description

[Technical Field]

[0001] The present invention relates to a method for determining the concentration of a substance in a solution in real time, and more particularly to a method for accurately predicting the concentration of a substance to be analyzed by learning a real-time optical analysis spectrum of a process solution in which the concentration of the substance changes based on a concentration prediction algorithm. [Background technology]

[0002] To obtain chemical information about a substance in a solution, spectra can be obtained using optical analysis methods using spectrometers such as infrared spectroscopy (IR), Raman spectroscopy, and ultraviolet-visible (UV-vis) spectroscopy. Based on the obtained spectra, chemical information about the substance can be extracted and the concentration can be calculated.

[0003] In analyzing substance concentrations using conventional optical analytical methods, the mixture state of a wide variety of chemical substances makes it difficult to distinguish the chemical information of each substance according to its components. This requires a long time to distinguish and analyze these components, making it difficult to measure the concentration of the target analyte in real time. To address this issue, conventional techniques have predicted the analyte concentration by performing multivariate linear combinations using machine learning-based multi-regression analysis, polynomial regression analysis, partial least squares (PLS), or net analyte signal (NAS) algorithms. Such methods are disclosed in Patent Document 1. However, when the relationship between the analyte concentration and the spectrum acquired from a spectrometer is not linear, the accuracy of forming an algorithmic model and predicting the concentration is low. Furthermore, when the change in the spectrum according to the analyte concentration in a solution is slight, the learning rate of the model decreases.

[0004] Meanwhile, the concentration of an analyte can be predicted from its spectrum using methods such as decision trees, random forests, support vector machines, and deep learning, which are capable of performing nonlinear regression learning based on machine learning. However, these methods generate a prediction model by simply adjusting learning parameters related to the spectrum, and the accuracy of predicting the concentration of a target substance is low due to the wide variety of substance information contained in the spectrum.

[0005] To solve this problem, Patent Document 2 attempts to predict the concentration more accurately by transforming the spectrum using a background signal, a comparison of similarity intervals, and the like.

[0006] However, these prior art techniques predict concentrations by simply modifying spectra, and lack spectral preprocessing techniques for optimal concentration prediction. To address the above-mentioned problems, the present invention provides a method and system for predicting the concentration of an analyte, which generates a concentration prediction model by performing machine learning using the modified spectrum of the substance, and then modifies the spectrum of the analyte and inputs it into the generated concentration prediction model, thereby improving the accuracy of the concentration prediction model.

[0007] Related prior art includes the following patent documents. [Prior art documents] [Patent documents]

[0008] [Patent Document 1] Republic of Korea Publication Patent No. 2020-0018177 [Patent Document 2] Japanese Patent No. 6871195 Summary of the Invention [Problem to be solved by the invention]

[0009] The present invention has been devised to solve the above-mentioned problems, and its purpose is to provide a method for predicting the concentration of an analyte that can improve the accuracy of concentration prediction by adding an analyte spectrum transformation algorithm to a concentration prediction model. [Means for solving the problem]

[0010] In order to achieve the above-mentioned object, the present invention provides a method for generating a concentration prediction model for predicting the concentration of a target substance by machine learning, the method including: a training data generation step of generating, as training data, an optimally transformed spectrum obtained by transforming a basic spectrum of a substance whose concentration is known based on predetermined transformation conditions; and a concentration prediction model generation step of performing machine learning on the optimally transformed spectrum transformed based on the predetermined transformation conditions generated in the training data generation step, and the optimally transformed spectrum and the measured concentrations of the substances corresponding to each of them, to generate a concentration prediction model.

[0011] The predetermined transformation conditions may be optimal transformation conditions derived by a spectral transformation and combination step of transforming a basic spectrum based on two or more mutually different transformation conditions to generate a transformed spectrum, and an optimal transformation condition derivation step of deriving optimal transformation conditions from the transformed spectrum.

[0012] The optimal transformation condition derivation step may be characterized by calculating the standard deviation of the reference similarity of the transformed spectrum generated in the spectral transformation and combination step, the correlation coefficient with the measured concentration, and the multivariate ratio before and after the spectral transformation, and deriving the transformation condition that maximizes the sum of these as the optimal transformation condition.

[0013] In addition, to achieve the above-mentioned object, the present invention provides a computer recording medium on which a concentration prediction model generation algorithm is recorded, the computer recording medium including: a spectrum transformation module that transforms a basic spectrum based on two or more transformation conditions to generate a transformed spectrum corresponding to the basic spectrum; an optimal transformed spectrum calculation module that calculates the standard deviation of the reference similarity of the transformed spectrum based on each of the transformation conditions, the correlation coefficient with the actual concentration of the substance corresponding to each transformed spectrum, and the multivariate ratio before and after the spectral transformation, and calculates an optimal transformed spectrum transformed based on the transformation condition that maximizes the sum of these; and a machine learning module that generates a concentration prediction model by machine learning using the actual concentration of the substance corresponding to the optimal transformed spectrum as learning data.

[0014] Furthermore, in order to achieve the above-mentioned object, the present invention provides a concentration prediction method for predicting the concentration of an analyte, comprising: a predicted spectrum acquisition step of acquiring a predicted spectrum of the substance whose concentration is to be predicted; an optimally modified predicted spectrum generation step of transforming the acquired predicted spectrum under optimal transformation conditions to generate an optimally modified predicted spectrum; and a predicted concentration output step of inputting the optimally modified predicted spectrum into a concentration prediction model and outputting a predicted concentration value, wherein the optimal transformation conditions are transformation conditions derived by a spectrum transformation and combination step of transforming a basic spectrum of a predetermined substance whose actual measured concentration value is known based on two or more mutually different transformation conditions to generate a transformed spectrum, and an optimal transformation condition derivation step of deriving the optimal transformation conditions from the generated transformed spectrum, and wherein the concentration prediction model is a neural network or concentration prediction algorithm trained to calculate a predicted concentration value from the optimally modified spectrum by machine learning the optimally modified spectrum obtained by transforming the basic spectrum under the optimal transformation conditions and the corresponding actual measured concentration value.

[0015] The optimal transformation condition derivation step may be characterized by calculating the standard deviation of the reference similarity of the transformed spectrum generated in the spectral transformation and combination step, the correlation coefficient with the measured concentration, and the multivariate ratio before and after the spectral transformation, and deriving the transformation condition that maximizes the sum of these as the optimal transformation condition.

[0016] Furthermore, in order to achieve the above-mentioned object, the present invention provides a concentration prediction system for predicting the concentration of an analyte, comprising: a spectral data acquisition unit that acquires a basic spectrum of a substance for generating learning data, or a predicted spectrum, which is the spectrum of a substance whose concentration is to be measured; a substance concentration data acquisition unit that acquires already acquired concentration data of the substance and the actual measured concentration of the substance to be measured using a known concentration measuring device; a calculation and control unit that generates modified spectra of the basic spectrum and the predicted spectrum, calculates optimal modification conditions, performs machine learning, and generates a concentration prediction model; and a memory that stores the basic spectrum of the analyte acquired by the spectral data acquisition unit, the actual measured substance concentration value acquired by the substance concentration data acquisition unit, and the concentration prediction model generated by the calculation and control unit.

[0017] The calculation and control device may be characterized in that it transforms and combines the spectra acquired by the spectral data acquisition unit based on a spectrum transformation algorithm, derives from the transformed spectra a transformed spectrum that indicates the accuracy of the optimal concentration prediction as an optimal transformed spectrum, performs machine learning using the optimal transformed spectrum and the actual measured concentration of the corresponding substance as learning data to generate a concentration prediction model, and inputs the predicted spectrum into the generated concentration prediction model to predict the concentration.

[0018] The memory may include a spectrum transformation and combination module that transforms the spectrum acquired by the spectrum acquisition unit based on transformation conditions; an optimal spectrum derivation module that acquires an optimal spectrum from the transformed spectra that indicates the accuracy of optimal concentration prediction and derives optimal transformation conditions; a machine learning module that uses the optimal transformed spectrum and the measured substance concentrations as learning data to advance machine learning and generate a concentration prediction model; and a concentration prediction module that calculates a concentration prediction value using the predicted spectrum as input data.

[0019] Furthermore, to achieve the above-mentioned object, the present invention provides a computer-readable recording medium having recorded thereon a concentration prediction algorithm, the computer-readable recording medium comprising: a spectrum transformation module that transforms a predicted spectrum of a substance whose concentration is to be predicted to generate a transformed spectrum; an optimal spectrum derivation module that calls the spectrum transformation module and calculates an optimally transformed predicted spectrum by transforming the predicted spectrum based on optimal transformation conditions; and a concentration prediction module that receives the optimally transformed predicted spectrum and calculates a predicted concentration value, wherein the concentration prediction module is generated by machine learning based on the spectrum transformed based on the optimal transformation conditions and the actual measured concentration of the substance corresponding to the spectrum.

[0020] Furthermore, in order to achieve the above-mentioned object, the present invention provides a concentration measuring device for predicting the concentration of an analyte, characterized in that it comprises a spectral data acquisition unit that acquires a predicted spectrum, which is the spectrum of the analyte, a memory device equipped with a spectrum transformation module that transforms the predicted spectrum based on optimal transformation conditions and a concentration prediction module that calculates a predicted concentration value from the optimally transformed predicted spectrum obtained by transforming the predicted spectrum based on the optimal transformation conditions, and an arithmetic and control device that reads the concentration prediction module equipped in the memory device and controls it to calculate a predicted concentration value from the predicted spectrum whose concentration is to be predicted.

[0021] The concentration measuring instrument may be provided with a data connection unit for connecting an external memory device and receiving input and output of data from an external device, and the memory device may be connected to the data connection unit in a removable form.

[0022] The concentration measuring instrument may include a data connection unit for connecting an external memory device and receiving input and output data from the external device, and the memory device may receive and store from the external device via the data connection unit a spectrum transformation module that transforms the predicted spectrum based on optimal transformation conditions and a concentration prediction module that calculates a predicted concentration value from the optimally transformed predicted spectrum obtained by transforming the predicted spectrum based on the optimal transformation conditions.

[0023] The data connection unit may be configured integrally with the spectral data acquisition unit. [Effects of the Invention]

[0024] The present invention can improve the accuracy of a concentration prediction model by performing machine learning on a modified spectrum of a substance to generate a concentration prediction model, and then modifying the spectrum of the analyte and inputting it into the generated concentration prediction model.

[0025] Furthermore, the present invention derives the optimal transformation method (transformation conditions) from among various transformation methods when transforming a spectrum, generates a concentration prediction model using a transformed spectrum transformed using the derived optimal transformation method (transformation conditions), and uses this to predict concentrations, thereby improving the accuracy of predicting substance concentrations. [Brief explanation of the drawings]

[0026] [Figure 1] 1 is a diagram illustrating a concentration prediction system and apparatus for predicting the concentration of an analyte of the present invention. [Figure 2] 1 is a block diagram of a concentration measuring device according to an embodiment of the present invention. [Figure 3]1 is a diagram illustrating a spectrum transformation method using an inverse transformation method according to an embodiment of the present invention. [Figure 4] 1 is a diagram illustrating a spectral transformation method using a filtering transformation method according to an embodiment of the present invention. [Figure 5] 1 is a spectrum modification method using a baseline removal modification method according to an embodiment of the present invention. [Figure 6] 1 is a diagram illustrating a spectrum transformation method using an interval averaging transformation method according to an embodiment of the present invention. [Figure 7] 1 is a diagram illustrating a spectrum transformation method using a section removal transformation method according to an embodiment of the present invention. [Figure 8] FIG. 1 is a flow chart showing a concentration prediction method of the present invention for generating a concentration prediction model and predicting the concentration of an analyte. [Figure 9] 10 is a graph showing the mean square error based on the optimal spectrum transformation conditions of the present invention. [Figure 10] 1 is a graph showing the concentration prediction results of a conventional concentration prediction method. [Figure 11] 1 is a graph showing the concentration prediction results of the concentration prediction method of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0027]

[0023] The following detailed description of the preferred embodiments of the present invention will be given with reference to the accompanying drawings so that those skilled in the art can easily implement the present invention. However, the present invention may be embodied in various different forms and is not limited to the preferred embodiments described herein. In the drawings, parts that are not relevant to the description are omitted for clarity, and similar parts are designated by similar reference numerals throughout the specification.

[0028] The present invention will be described in detail below with reference to the accompanying drawings.

[0029] The present invention relates to a method for predicting the concentration of an analyte.

[0030] More particularly, the present invention relates to a substance concentration prediction system and method that improves accuracy by adding a spectral transformation algorithm to a substance concentration prediction model.

[0031] 1. Generation of concentration prediction model and concentration prediction method of the present invention

[0032] The concentration prediction model of the present invention is generated by machine learning, in which the concentration of a substance whose concentration is known and obtained from a device capable of obtaining the spectrum of the substance, such as a known spectrometer, or whose concentration can be measured experimentally, is transformed, and the transformed spectrum is learned to predict the concentration value.

[0033] 1.1. Training data generation step

[0034] The present invention uses, as training data, the transformed spectrum obtained by transforming the spectrum of a substance whose concentration is known as described above and the measured concentration of the substance. Furthermore, to generate an optimal prediction model, the present invention determines optimal spectrum transformation conditions that generate an optimal prediction model from among various spectrum transformation methods and spectrum transformation conditions obtained by combining these methods, and uses, as training data, the transformed spectral data transformed under the optimal spectrum transformation conditions and the measured concentration of the substance.

[0035] 1.1.1. Step of Acquiring Fundamental Spectrum Si Data (S10)

[0036] In the present invention, the i fundamental spectra to be used as training data are denoted as Si. In the present invention, the fundamental spectra refer to the initial spectral data acquired using a known spectrometer or the like. In the present invention, the i acquired spectra S are denoted as S1, S2, ..., Si.

[0037] 1.1.2. Measured substance concentration LC i Data acquisition step (S20)

[0038] The measured concentration data LCi of each substance corresponding to each of the fundamental spectra described above is obtained. When a fundamental spectrum Si is obtained for a substance whose concentration is already known, the concentration data LCi of each substance may be a value obtained by experimentally measuring the concentration of the substance whose fundamental spectrum has been obtained using a known concentration measuring device linked to the concentration measurement system of the present invention, using the known substance concentration data.

[0039] 1.1.3. Step of obtaining the optimal deformation spectrum of MoSi (S30)

[0040] The modified spectrum used as training data in the present invention is a spectrum obtained by modifying the above-mentioned basic spectrum based on the modification conditions described below.

[0041] In the present invention, for the sake of explanation, k types of spectral transformation conditions can be set according to the transformation conditions and their combinations described below, and the first, second, ..., kth transformation conditions are represented as M1, M2, ..., Mk, respectively, and the optimal transformation condition for generating the optimal concentration prediction model is represented as Mo.

[0042] It should be obvious to ordinary engineers that the transformed spectrum under each transformation condition can be expressed as M1S1, M1S2, ... M1Si when the basic spectrum (Si, i = 1, 2, ..., n) is transformed under transformation condition M1, and the transformed spectrum under transformation condition Mk can be expressed as MkSi (k = 1, 2, ..., n, i = 1, 2, ..., M).

[0043] In the present invention, of these modified spectra MkSi, the modified spectrum MoSi, which is modified under the optimum modification condition Mo that indicates the optimum accuracy of concentration prediction, is used as training data.

[0044] (1) Spectral transformation and combination step

[0045] This is a step in which the basic spectrum Si is transformed based on a wide variety of transformation conditions and combined to derive the optimal transformation conditions.

[0046] Below are examples of five basic spectral modification conditions used in the present invention. In the present invention, as long as any one of the following five modification methods can be used, at least two or more of these methods can be selectively combined and used as spectral modification conditions. However, the present invention is not limited to these five methods and their combinations, and spectral modification conditions and their combinations that were known before the filing date of the present invention can also be used as spectral modification conditions.

[0047] (i) Deformation condition 1: Inverse deformation

[0048] FIG. 3 shows a spectrum transformation method using the inverse transformation method according to an embodiment of the present invention. The inverse transformation means transforming the data values of the acquired spectrum into the inverse numbers. For example, the intensity of the spectrum is transformed into the inverse number I x Then, the intensity of the spectrum represented by the inverse transformation can be transformed into: where R is a variable that removes large deviations from the spectrum, so that the deviation of the acquired inverse transformed spectrum data does not appear larger than a certain level.

[0049] (ii) Transformation condition 2: Filtering transformation

[0050] 4 illustrates a spectral modification method using a filtering modification method according to an embodiment of the present invention. Filtering modification refers to a modification that selectively blocks a portion of an acquired spectrum and transmits the remainder. By filtering out spectral peaks that are unrelated to the analyte, the relationship between the actual analyte and the acquired spectrum can be improved.

[0051] (iii) Transformation condition 3: Baseline removal transformation

[0052] FIG. 5 illustrates a spectrum modification method using baseline removal modification according to an embodiment of the present invention. When a spectrum is acquired using a spectrometer, the spectral data may not have an ideal baseline due to noise caused by system characteristics and variations in the measured substance sample. Therefore, removing the baseline to remove the effects of baseline shift from the spectral data can improve the relationship between the concentration of the measured substance and the spectrum. Because baseline removal modification can increase the similarity of data between spectra, its use can be determined depending on the type of measured substance or the analytical method.

[0053] (iv) Transformation condition 4: Interval averaging transformation

[0054] FIG. 6 illustrates a spectrum modification method using interval averaging according to an embodiment of the present invention. Interval averaging refers to a modification that averages data values within a predetermined interval of an acquired spectrum. In this case, not only the spectral data within the predetermined interval but also all variables are averaged. Interval averaging can also remove unwanted peaks due to noise in the spectrometer system, improving the relationship between the analyte concentration and the spectrum.

[0055] (v) Transformation condition 5: Section removal transformation

[0056] 7 illustrates a spectrum transformation method using the interval elimination transformation method according to an embodiment of the present invention. The interval elimination transformation is a transformation that removes data values in a predetermined interval from an acquired spectrum. When peaks or data in a spectrum contain peaks or data unrelated to the analyte to be measured, the interval elimination transformation can improve the relationship between the analyte concentration and the spectrum over the entire interval.

[0057] (2) Derivation step of the optimal deformation condition Mo

[0058] This is the step of deriving the optimum deformation conditions according to the procedure described below.

[0059] The inventors of the present invention discovered that an optimal prediction model can be generated when the optimal deformation conditions are set to maximize the function values of the standard deviation of the reference similarity (Ak), the correlation coefficient with concentration (Bk), and the multivariate ratio (Ck) of the deformed spectra (M1Si, M2S2, ..., MkSi) deformed under each deformation condition, and applied this to the present invention.

[0060] In particular, the optimal transformation conditions were those that maximized the value of Fk(Ak,Bk,Ck)=aAk+bBk+cCk, which is expressed as a linear function of the Ak, Bk, and Ck values. Here, when transformation is performed under transformation condition M1, the linear function is expressed as F1(A1,B1,C1), where a, b, and c are constant values set by the user and can be omitted.

[0061] That is, in the present invention, the optimal transformation condition means deriving the k that maximizes the Fk(Ak, Bk, Ck), and the Fo(Ao, Bo, Co) = Ao + Ao + Co value for a spectrum transformed under the optimal transformation condition Mo is the maximum value among the Fk (k = 1, 2, ..., n) values.

[0062] Preferably, in the present invention, the optimal transformation condition refers to deriving k such that Fk(Ak, Bk, Ck) is maximized, and the value of Fo(Ao, Bo, Co) = aAo + bAo + cCo for a spectrum transformed under the optimal transformation condition Mo is the maximum value among the Fk (k = 1, 2, ..., n) values. In this case, the constants a, b, and c may be values selected from the ranges of 6.5 to 7.7, 0.5 to 2.0, and 2.5 to 4.0, and more preferably, a, b, and c may be selected as 6.69, 1.52, and 3.59, respectively.

[0063] For example, in the spectrum optimal transformation condition derivation unit of the present invention, i spectra Si acquired as learning data are transformed using the above-mentioned transformation conditions 1 to 10, and the Fk value is calculated for each of the k types of transformed spectra that are combined, and the kth transformation condition Mk that shows the largest Fk is derived as the optimal transformation condition Mo.

[0064] Below, we will explain the values of the standard deviation of the reference similarity of the deformed spectrum (Ak), the correlation coefficient with concentration (Bk), and the multivariate ratio (Ck).

[0065] (i) Standard deviation of the reference similarity (A)

[0066] In order to derive the optimum spectrum transformation conditions, the standard deviation of the reference similarity between the reference spectrum and each transformed spectrum transformed under different transformation conditions according to the spectrum transformation methods or combinations thereof is calculated.

[0067] Here, the reference spectrum is a linear spectrum in which the intensity in all wavelength bands is K (a constant other than 0). Therefore, the standard deviation of the reference similarity is calculated by calculating the standard deviation of the reference similarity between all the transformed spectra and the reference spectrum. For example, if there are 10 samples of the basic spectrum before transformation (S1, S2, ..., S 10 ) and there are 10 types of transformation combination methods (M1, M2, ..., M 10 ) and transform each of them, we get 100 transformed spectra (M1S 1, …,M1S 10 ,M2S1,…,M2S 10 ,…M 10 S 10 ) is generated, and 100 reference similarities can be calculated based on the reference spectrum.

[0068] A reference similarity of the transformed spectrum based on each of the transformation conditions is calculated.

[0069] For example, the standard deviation (A1) of the reference similarity is calculated for the transformed spectrum M1Si transformed based on the transformation condition M1, and the standard deviation (Ak) of the reference similarity is calculated for the transformed spectrum M2Si, ..., MkSi transformed based on the transformation conditions M2, ..., Mk in the same manner.

[0070] The standard deviation of the reference similarity is calculated by calculating the cosine similarity of the transformed spectrum under different transformation conditions compared to the reference spectrum. The cosine similarity indicates the similarity that can be calculated using the cosine angle between a linear reference spectrum and a transformed spectrum, i.e., the two spectra, and the closer the value is to 1, the higher the similarity can be determined. The reference similarity is calculated according to the following Equation 1.

[0071]

number

[0072] In the above Equation 1, α represents the combined intensity of the combined transformed spectra, and β represents the combined intensity of the reference spectra, which is an adjustable constant that has a constant value for all multivariates.

[0073] The reference similarity value between the modified spectrum and the reference spectrum based on each modification condition is calculated using Equation 1 above, and the standard deviation of the reference similarity is the result of calculating the standard deviation between the reference similarities (Ak). In another embodiment, when the number of modified spectra is greater than a predetermined standard, the result of dividing the standard deviation between the similarities by the average similarity is

number

[0074] Therefore, the standard deviation A of the reference similarity calculated in the present invention is a value that allows the degree of difference between each of the deformed spectra to be analogized, and enables sensitive recognition of the shape difference between the deformed spectra.

[0075] Deformed spectrum M1S deformed under deformation condition M1 i The standard deviation of the reference similarity for the deformation spectrum M1S1, M1S2, ..., M1S i of

number

[0076] It is calculated as:

[0077] By this method, the deformation condition M k The deformed spectrum M k S i Calculate the standard deviation (Ak) of the reference similarity for

[0078] (ii) Correlation coefficient with substance concentration (B)

[0079] The relationship with substance concentration refers to a normalized value (B) obtained by normalizing the covariance value, which indicates the relationship between the reference similarity of the spectrum and the corresponding actual substance concentration (measured concentration LCi), to a maximum value of 1. The value obtained by normalizing the covariance value to a maximum value of 1 is the same as the correlation coefficient of covariance, which is characterized by having a range from -1 to +1, so the resulting value is the same even when calculated using the correlation coefficient of covariance. In other words, the calculation of the relationship with the target substance concentration can estimate the relationship between the shape of the deformed spectrum and the corresponding substance concentration using B.

[0080] The normalized value of covariance or the correlation coefficient of covariance is calculated according to the following Equation 2, and the resulting value is in the range of −1 or more and +1 or less.

[0081]

number

[0082] In the above Equation 2, X represents the reference similarity and Y represents the actual substance concentration of the analyte.

[0083] Furthermore, Cov(X,Y) indicates the covariance value between the reference similarity (X) and the corresponding concentration of the analyte (Y), and Var(X) and Var(Y) indicate the variance values between the reference similarity (X) and the corresponding concentration of the analyte (Y).

[0084] In the above, the actual concentration of a substance refers to the actual concentration of a substance measured by an experiment or other method, or the actual concentration of a known substance. This process reflects the relationship between the actual concentration of a substance and the deformation spectrum as a learning element.

[0085] The LCi value, which is the actual concentration value of the spectrum-corresponding substance previously acquired as learning data, is input as the Y value in the above Equation 2. The previously calculated reference similarity value is input as the X value.

[0086] The normalized value (Bk) of the covariance value calculated according to the above formula is calculated according to the combination of the modified spectrum. For example, in the modified combination method M k A reference similarity is calculated for the transformed spectrum M1Si, and the normalized value (Bk) of the covariance value is calculated based on the actual concentration value LCi of the corresponding substance.

[0087] (iii) Multivariate ratio (C) before and after spectral transformation

[0088] In the present invention, a multivariate ratio (C) is calculated as another variable of the optimal spectral transformation condition. The multivariate ratio is

number

[0089] The number of multivariates in the present invention refers to the number of data acquired in a predetermined wavelength band corresponding to the X-axis of the acquired spectrum. For example, if a spectrometer acquires a spectrum in a wavelength band from 1000 nm to 3000 nm in 1 nm increments, the number of multivariates at this time is (3000-1000), which is 2000. If the spectrum from 1000 to 1500 nm is removed using interval removal transformation among the transformation methods, the number of multivariates after transformation is (3000-1500), which is 1500 in total. As another example, when interval averaging transformation is performed, the number of multivariates is reduced to the number of set intervals.

[0090] Using this method, the number of multivariates after spectral transformation can be calculated compared to the number of multivariates before spectral transformation to obtain the C value. In this case, C can be used to estimate the degrees of freedom for the number of variables after spectral transformation when learning spectral transformation using an ML model.

[0091] The multivariate ratio (Ck) calculated according to the above formula varies depending on the data values before and after spectral transformation, and therefore indicates a representative value of the degree of freedom for the data values due to spectral transformation.

[0092] (3) Calculation of the optimal deformation spectrum MoSi data

[0093] After determining Mo, which is the transformation condition that maximizes Fk(Ak, Bk, Ck) for the i*k transformed spectra MkSi using the above-mentioned method, the basic spectrum Si is transformed using the transformation condition Mo to calculate the optimal transformed spectrum MoSi data. The optimal transformed spectrum MoSi data may be the transformed spectrum MoSi data for k=0 previously generated in the calculation process of Fk(Ak, Bk, Ck) as is.

[0094] 1.2. Machine Learning and Concentration Prediction Model Generation Step (S40)

[0095] In the present invention, a concentration prediction model is generated by machine learning using optimally transformed spectra MoSi, which are obtained by transforming the above-mentioned fundamental spectrum Si under optimal transformation conditions Mo, and the measured concentrations LCi of the corresponding substances as training data.

[0096] The neural network and learning model used to train the concentration prediction model generated in the present invention are not limited to one specific method, and any neural network or learning model that generates a concentration prediction model from known spectral data may be used.

[0097] In this way, the concentration prediction model generated by learning using the optimal modified spectrum MoSi is used to predict concentration in the concentration prediction step described later.

[0098] 1.3. Concentration prediction step

[0099] The generated concentration prediction model is used to predict the concentration of a new substance. The concentration prediction includes a step of acquiring a predicted spectrum PS, a step of generating an optimally modified predicted spectrum P.MoS by transforming the acquired predicted spectrum under the above-mentioned optimal transformation conditions, and a step of outputting a predicted concentration by inputting the optimally modified predicted spectrum P.MoS data into the generated concentration prediction model and outputting a predicted concentration value PLC.

[0100] (1) Step of obtaining predicted spectrum PS (P10)

[0101] This is a step in which the spectrum of the substance whose concentration is to be predicted is acquired by the spectrum acquisition unit.

[0102] (2) Generation step of the optimal deformation prediction spectrum P.MoS (P20)

[0103] This is the step of transforming the predicted spectrum under the above-mentioned optimal transformation conditions to generate an optimally transformed predicted spectrum.

[0104] (3) Predicted concentration PLC output step (P30)

[0105] This is the step of inputting the above-mentioned optimal deformation predicted spectrum into a concentration prediction model generated by machine learning using the above-mentioned optimal deformation spectrum set, and obtaining the concentration predicted value PLC, which is the output.

[0106] <Verification of the performance of the prediction model based on the detection of the optimal deformation conditions of the present invention>

[0107] Figure 9 shows the mean square error (MSE) based on the spectral transformation conditions. The graph shows the log(MSE average) based on F(A,B,C). -1 where MSE stands for Mean Square Error. In Figure 9, the higher F(A, B, C), the higher the log(MSE average). -1 It can be seen that the mean square error (MSE) tends to be higher. The lower the mean square error, the smaller the difference in concentration prediction values using the ML model, and therefore the higher the accuracy of the generated ML model. In other words, the lower the MSE average, the more advantageous it is to generate an ML model with high accuracy using spectral transformations.

[0108] 10 and 11 are graphs showing the concentration prediction results of the conventional concentration prediction method and the concentration prediction results of the concentration prediction method of the present invention, respectively. Fig. 10 shows the verification results for a model trained using a deep neural network (DNN) without spectral transformation according to the conventional technique, and Fig. 11 shows the verification results for a model trained using a DNN after optimal spectral transformation according to the present invention.

[0109] The data used in generating and validating the learning model of the present invention was used to validate predictions and results using a model that predicts the concentration of glucose used as a substrate from the real-time spectrum of the process solution in the incubator. In this validation, the value of Fo(Ao,Bo,Co) = aAo + bAo + cCo for the spectrum transformed under the optimal transformation condition Mo was the maximum value among the Fk (k = 1, 2, ..., n) values, where a, b, and c were applied as 6.69, 1.52, and 3.59, respectively.

[0110] R 2 The line drawn with indicates the regression coefficient of determination, and the larger its value, the better the predictive results of the model.

[0111] To evaluate the trained and generated concentration prediction ML model, data is excluded in advance from use in training and generating the concentration prediction ML model, and the model's performance is tested using this excluded data (called validation data). In the graph, the fimse (green dots) represent the results of evaluating the mean squared error (MSE) based on validation data not used in training and generating the ML model. Also, the mmse (black dots) represent the results of evaluating the mean squared error (MSE) based on training data used in training and generating the ML model, and the pmse (red dots) represent the results of evaluating the mean squared error based on input data not used in training and generating the ML model.

[0112] The accuracy of the concentration prediction of the ML model generated using the optimal spectrum transformation process of the present invention shown in FIG. 11 is higher than that of the conventional technique shown in FIG. 2 It can be seen that the values are high, and that the PMSE and FIMSE point data are also more precisely and accurately distributed on the predicted validation line.

[0113] 2. Substance concentration prediction system 100 according to the present invention

[0114] FIG. 1 is a diagram showing a substance concentration prediction system 100 according to the present invention.

[0115] 1, the present invention is based on a spectrometer for an optical analysis method, in which the spectrometer is used to obtain spectra of a substance whose concentration is known and a substance whose concentration is to be predicted, and the obtained spectra are analyzed to output a concentration value of the substance. The detected spectra and the concentration value of the analyte are applied as learning data for a concentration prediction model, and optimal transformation conditions for the spectrum are derived using the concentration prediction model to more accurately predict the concentration. The concentration prediction model transforms the spectrum according to the optimal transformation conditions to predict the concentration of the analyte.

[0116] In FIG. 1, the data flow indicated by the solid lines represents the data flow used to generate the concentration prediction model, and the data flow indicated by the dotted lines represents the data flow associated with the procedure for predicting the concentration of a substance using the generated concentration prediction model.

[0117] The substance concentration prediction system 100 of the present invention comprises a spectrum data acquisition unit 10, a concentration data acquisition unit 20, a calculation unit 30, and a memory 40. Each component will be described below.

[0118] 2.1. Spectral data acquisition unit 10

[0119] The spectrum acquisition unit may be a known spectrometer or spectrometer for optical analysis. The spectrum of the analyte whose concentration is to be measured is acquired using the spectrometer, and the acquired spectrum can be used to calculate the concentration of the analyte by using an analysis method appropriate for the type of spectrometer.

[0120] The spectrum acquisition unit of the present invention acquires a fundamental spectrum Si for use as training data and a predicted spectrum PSi for predicting concentrations. The spectrum acquisition unit 10 can acquire fundamental spectra or predicted spectra from an external device such as a spectrometer, an external spectrum generator, a memory 40, or the like, and may include a data connection unit 11 for connecting data.

[0121] (1) Acquisition of the fundamental spectrum Si

[0122] The spectral data acquisition unit 10 acquires fundamental spectra from a spectrometer of a substance with a known concentration, and uses the acquired fundamental spectrum set as training data. In the present invention, the acquired i spectra S are denoted as S1, S2, ..., S i The basic spectrum is the basis for generating a transformed spectrum MSi to be used as training data in the machine learning module 43, which will be described later.

[0123] The fundamental spectrum S i may be stored in advance in the memory in the form of already acquired spectrum data, in which case the spectrum acquisition unit acquires the fundamental spectrum S i from the memory.

[0124] (2) Obtaining the predicted spectrum PSi

[0125] The spectral data acquisition unit 10 also acquires a predicted spectrum PSi of the substance to be used as input for the generated concentration prediction model to predict the concentration. The acquired predicted spectrum data is transformed into a predicted optimally transformed spectrum PMoSi based on the optimal transformation conditions, and input into the concentration prediction model generated according to the above-mentioned procedure to be used for predicting the concentration of the substance.

[0126] 2.2. Substance concentration LCi data acquisition unit 20

[0127] This is a component that acquires the concentration data LCi of substances corresponding to the above-mentioned fundamental spectra. In the case of known substance concentrations, the concentration data LCi is stored in a memory and can be acquired by reading the memory in the substance concentration LCi data acquisition unit 20. Alternatively, the substance concentration data LCi corresponding to the fundamental spectra is acquired by measuring using a known concentration measuring device (not shown) that is linked to the concentration measurement system of the present invention or by experimentally measuring.

[0128] The concentration data acquisition unit 20 can acquire known or experimentally measured concentration data from an external concentration measuring device, an external experimental data storage, memory 40, etc., and may be equipped with a data connection unit 21 for connecting the data.

[0129] 2.3. Calculation and Control Unit 30

[0130] The calculation and control device 30 of the present invention may be a single CPU or microprocessor, or a terminal device, computer device, computer server, or the like that includes these.

[0131] The calculation and control device 30 controls the overall operation of the system, acquires and stores data, performs machine learning and generates concentration prediction models, and manages the concentration prediction models generated as a result of the machine learning.

[0132] The calculation and control device 30 also controls the transformation of the spectrum in the present invention, the derivation of the optimal transformation conditions using the transformed spectrum, the generation of the transformed spectrum based on the optimal transformation conditions, the learning of the concentration prediction model using the transformed spectrum, and the generation of the concentration prediction model.

[0133] The calculation and control device of the present invention can utilize a computer program stored in memory to control each function when performing these operations, and stores the generated concentration prediction model in the memory. When predicting the concentration, the concentration prediction model is read from the memory again to perform the prediction calculation. It will be easily understood by those skilled in the art that the computer programs for performing each of these functions are stored as separate modules in the memory device described below, and that when the calculation and control device is needed, these can be read from the memory to perform the function of each module.

[0134] Each operation performed by the calculation and control device of the present invention will be described below.

[0135] (1) Spectral transformation and combination

[0136] The fundamental spectrum Si acquired by the spectrum acquisition unit is transformed and combined to generate a transformed spectrum MkSi. This operation is as described above in Section 1.1.3. In performing this procedure, the calculation and control device can load a spectrum transformation and combination module or transformation algorithm stored in memory in the form of software / program.

[0137] In the present invention, the spectral transformation and combination procedure may be configured as a separate spectral transformation and combination unit to perform the spectral transformation and combination procedure based on a spectral transformation algorithm using a microprocessor or the like configured as a separate module in the arithmetic and control device.

[0138] (2) Derivation of the optimal deformation spectrum

[0139] The optimal deformation conditions are derived from the above-mentioned deformation spectrum MkSi. The specific procedure for deriving the optimal deformation conditions is as explained in Section 1.1.3., and can be performed by loading the optimal spectrum derivation module installed in memory in the form of software / program, which will be described later.

[0140] (i) The optimal transformed spectrum used as learning data to generate a concentration prediction model is calculated by determining optimal transformation conditions from the transformed spectrum for the basic spectrum S and transforming the basic spectrum S under the optimal transformation conditions, or by selecting a transformed spectrum that meets the optimal transformation conditions from among already generated transformed spectra. The derived optimal transformed spectrum is used as input data for machine learning, which will be described later.

[0141] The derived optimum deformation condition Mo may be stored in memory for future use.

[0142] (ii) When a concentration is to be predicted, the calculation and control device transforms the predicted spectrum PSi of the target substance using the previously calculated optimal transformation conditions Mo to generate an optimally transformed predicted spectrum MoPSi. The optimally transformed predicted spectrum MoPSi is input to the concentration prediction module 44, which calculates the predicted concentration value PLC.

[0143] In another embodiment of the present invention, the spectrum transformation and combination unit may be separately configured using a separate microprocessor or the like to perform the procedure for deriving the optimal transformed spectrum.

[0144] (3) Machine learning and generation of concentration prediction model

[0145] As described in Section 1.2 above, the control and calculation unit of the present invention performs machine learning using optimally transformed spectra MoSi obtained by transforming the fundamental spectrum Si under the optimal transformation conditions Mo and the corresponding measured concentrations LCi of substances as training data to generate a concentration prediction model. The generated concentration prediction model may be stored in memory in the form of a software program or algorithm.

[0146] (4) Concentration prediction

[0147] The previously generated optimal deformation prediction spectrum MoPSi is input to the generated concentration prediction model to calculate the concentration prediction value PLC.

[0148] 2.4. Memory Device 40

[0149] The memory device 40 in the concentration prediction system of the present invention can be used as a data storage for storing in advance concentration value data LCi of substances with known concentrations and their fundamental spectra Si, which are used as training data, and is equipped with the following computer software modules in the form of computer programs / software to perform the various functions of the system of the present invention. The memory device 40 also stores the optimal deformation conditions Mo calculated by the control and calculation unit.

[0150] The concentration prediction system of the present invention may be configured by incorporating the following software modules, but each module may be configured separately in a separate memory location. The memory device 40 in the present invention is not necessarily limited to a single physical memory location.

[0151] Furthermore, the memory of the present invention may be replaced with a memory device of a computer system, a server system, etc., or may be realized as a removable memory device having a separate CD, hard disk, USB, etc. that contains each program module or each software algorithm described below.

[0152] Each "module" described below may refer to a series of computer software algorithms configured to perform a given function.

[0153] (1) Spectral Transformation and Combination Module 41

[0154] The memory of the present invention stores the various transformation methods and computer software modules or transformation algorithms for performing the (1) spectral transformation and combination step described in the step of obtaining the optimal transformed spectrum in Section 1.1.3 above.

[0155] (2) Optimal spectrum derivation module 42

[0156] The memory of the present invention is equipped with a computer software module that performs the (2) optimal transformation condition derivation step described in the optimal transformation spectrum acquisition step in Section 1.1.3 above, and calls the spectral transformation and combination module to calculate the optimal transformation spectrum based on the optimal transformation conditions.

[0157] (3) Machine Learning Module 43

[0158] The memory of the present invention is equipped with a computer software module that functions as a machine learning module that calculates the (3) optimal deformed spectrum data described in the optimal deformed spectrum acquisition step in Section 1.1.3 above, uses this as learning data together with concentration value data LCi of substances with known concentrations, and proceeds with the machine learning process described in Section 1.2 above to generate a concentration prediction model.

[0159] (4) Concentration Prediction Module 44

[0160] The computer is equipped with a computer software module that executes the concentration prediction model generated by the machine learning module 43.

[0161] When attempting to predict the concentration of a given substance, the calculation and control device 30 inputs the predicted spectrum PSi of the substance whose concentration is to be predicted as input data, and inputs its modified spectrum, the optimal modified predicted spectrum PMoSi, into the concentration prediction module 44 of the prediction memory, and the concentration prediction module 44 calculates the concentration predicted value PLCi.

[0162] The concentration measurement system of the present invention may be configured to change and set optimal transformation conditions to correspond to optimal transformation conditions that are set differently depending on the type of substance to be measured. The calculation and control device 30 receives the input of predetermined optimal transformation conditions determined by a selection input displayed on a screen display unit (not shown) of the system or a selection input input from a physical switch (not shown), calls up a corresponding spectrum transformation module from memory 40 to transform the spectrum, and inputs the transformed spectrum into a concentration prediction model called up from memory 40 to output a predicted concentration value. In this case, it will be obvious to those skilled in the art that concentration prediction modules 44 corresponding to multiple concentration prediction models each corresponding to a specific optimal transformation condition can be generated and stored in memory 40 according to the concentration prediction model generation method described above.

[0163] 3. Concentration measuring device according to the present invention

[0164] Taking FIG. 2 as an example, a concentration measuring device equipped with a concentration prediction module 44 created based on the technical concept of the present invention will be described.

[0165] The concentration measuring device of the present invention is equipped with the spectral data acquisition unit 10, the calculation and control device 30, and the memory device 40 described in the previous two sections. Each component can be configured differently from the concentration measuring system described in the previous two sections, which will be described in detail below.

[0166] (1) Spectral data acquisition unit 10

[0167] The spectral data acquisition unit 10 of the concentration measuring instrument of the present invention acquires predicted spectral data PSi of a target substance. The input includes data output by known spectrometers, spectroscopes, optical sensors, and the like that output spectral data of a substance whose concentration is to be measured contact or non-contact.

[0168] (2) Memory device 40

[0169] The memory device 40 of the concentration measuring device of the present invention is equipped with the (1) spectral transformation and combination module mounted on the memory device 40 described in Section 2.4, the derived optimal transformation condition Mo, and the (4) concentration prediction module 44 based on the already learned optimal transformed spectrum.

[0170] Depending on the application, the memory device 40 of the concentration measuring instrument may be configured to include only a spectral transformation module based on the already calculated optimal transformation conditions among the spectral transformation and combination modules. In this case, the concentration prediction module 44 to be installed is also a concentration prediction module 44 that has already been trained using the optimally transformed spectrum as training data.

[0171] (3) Calculation and Control Device 30

[0172] The calculation and control device 30 of the concentration measuring instrument of the present invention utilizes the optimal transformation conditions, the spectrum transformation and combination module, and the concentration prediction module 44 installed in the memory device 40 to calculate the concentration prediction value PLCi from the acquired predicted spectrum PSi data of the target substance.

[0173] The concentration measuring device of the present invention is configured to change and set optimal transformation conditions to correspond to optimal transformation conditions that are set differently depending on the type of substance to be measured. The calculation and control device 30 receives predetermined optimal transformation conditions by accepting a selection input displayed on a screen display (not shown) or a selection input entered through a physical switch (not shown), and then retrieves a corresponding spectrum transformation module from the memory device 40 to transform the spectrum. The resulting spectrum is then input into a concentration prediction model retrieved from the memory device 40, and a predicted concentration value is output. It will be obvious to those skilled in the art that in this case, concentration prediction modules 44 corresponding to multiple concentration prediction models each corresponding to a specific optimal transformation condition may be generated and stored in the memory device 40 using the above-described method, and the device may be controlled to retrieve the concentration prediction module 44 appropriate for the corresponding substance to calculate the predicted concentration. Alternatively, the concentration prediction module 44 may be a general-purpose spectral concentration prediction model rather than one generated by learning a specific transformed spectrum.

[0174] (4) Data connection unit 21

[0175] On the other hand, the concentration measuring device of the present invention may further include a data connection section for connecting an external memory device 40 and receiving input and output of data from an external device.

[0176] The data connection unit is configured integrally with the spectrum data acquisition unit 10, and can also receive input of data other than spectrum data from an external device.

[0177] The spectral transformation and combination module 41, the optimum transformation conditions, and the concentration prediction model that have already been generated can be received via the data connection unit and stored in the memory device 40. With this configuration, the optimum transformation conditions, the spectral transformation and combination module 41, and the concentration prediction model that are set differently depending on the characteristics of the substance whose concentration is to be measured can be stored in the memory device 40 of the concentration measuring device and applied to the measurement of the concentration.

[0178] The data connection unit may be configured to be able to connect an external, removable memory device 40. In this case, the spectral transformation and combination module 41, the optimal transformation conditions, and the already generated concentration prediction model are stored in the external memory device 40, which is connected in addition to or in place of the memory device 40, and the calculation and control device 30 is configured to read them and use them for concentration prediction. [Explanation of symbols]

[0179] The names of the symbols used in the specification and drawings of the present invention are as follows:

[0180] 100: Concentration prediction system 10: Spectral data acquisition unit 20: Substance concentration data acquisition unit 30: Calculation and control device 40: Memory device 41: Spectral transformation and combination module 42: Optimal spectrum derivation module 43: Machine Learning Module 44: Concentration prediction module

Claims

1. A method for generating a concentration prediction model for predicting a concentration of a target substance by machine learning, comprising: a training data generating step of generating, as training data, an optimally transformed spectrum obtained by transforming a basic spectrum of a substance having a known concentration based on predetermined transformation conditions; a concentration prediction model generation step of generating a concentration prediction model by machine learning the optimally deformed spectra, which are deformed based on predetermined deformation conditions and which are generated in the training data generation step, and the measured concentrations of substances corresponding to the optimally deformed spectra; Including, The predetermined deformation condition is: a spectral transformation and combination step of transforming the basic spectrum based on two or more mutually different transformation conditions to generate a transformed spectrum; an optimal deformation condition derivation step of deriving optimal deformation conditions from the deformation spectrum; The optimal deformation condition is derived by The optimal deformation condition deriving step includes: calculating the standard deviation of the reference similarity of the transformed spectrum generated in the spectrum transformation and combination step, the correlation coefficient with the measured concentration, and the multivariate ratio before and after the spectrum transformation, and deriving the transformation conditions that maximize the sum of these as the optimal transformation conditions; Method for generating concentration prediction models.

2. The optimal deformation condition deriving step includes: the standard deviation (Ak) of the reference similarity, which is the standard deviation between the reference similarities of the transformed spectra for each transformation condition generated in the spectral transformation and combination step, the correlation coefficient (Bk) with the measured concentration, and the multivariate ratio (Ck) before and after the spectral transformation are calculated, and the transformation condition that maximizes the value of a linear function of these, Fk(Ak, Bk, Ck) = aAk + bBk + cCk (a, b, c are constant values), is derived as the optimal transformation condition; 2. The method for generating a concentration prediction model according to claim 1, wherein the reference similarity for calculating the standard deviation of the reference similarity, the correlation coefficient with the measured concentration, and the multivariate ratio before and after the spectral transformation are calculated by the following Equations 1, 2, and 3, respectively. [Equation 6] (α is the combined magnitude combination of the transformed spectra, and β is the combined magnitude combination of the reference spectra.) [Equation 7] (X is the reference similarity, Y is the actual concentration of the analyte, Cov(X,Y) is the covariance between the reference similarity (X) and the corresponding concentration of the analyte (Y), and Var(X) and Var(Y) are the variances of the reference similarity (X) and the corresponding concentration of the analyte (Y).) [Equation 8] (The number of multivariates is the number of data points acquired in a given wavelength band corresponding to the X-axis of the acquired spectrum.)

3. A spectral transformation module that transforms a basic spectrum of a substance having a known concentration based on two or more transformation conditions to generate a transformed spectrum corresponding to the basic spectrum; an optimal modified spectrum calculation module that calculates the standard deviation of the reference similarity of the modified spectrum based on each of the modification conditions, the correlation coefficient with the measured concentration of the substance corresponding to each of the modified spectra, and the multivariate ratio before and after the spectral modification, and calculates an optimal modified spectrum modified based on the modification condition that maximizes the sum of these; a machine learning module that performs machine learning using the measured concentrations of the substances corresponding to the optimal deformation spectrum as learning data to generate a concentration prediction model; A computer-readable storage medium having recorded thereon an algorithm for generating a concentration prediction model, comprising:

4. 1. A concentration prediction method for predicting a concentration of an analyte, comprising: a predicted spectrum acquisition step of acquiring a predicted spectrum of a substance whose concentration is to be predicted; an optimally modified predicted spectrum generating step of modifying the obtained predicted spectrum under optimal modification conditions to generate an optimally modified predicted spectrum; a predicted concentration output step of inputting the optimal deformation predicted spectrum into a concentration prediction model and outputting a predicted concentration value; comprising The optimum deformation condition is a spectrum transformation and combination step of transforming a basic spectrum of a predetermined substance whose actually measured concentration is known based on two or more different transformation conditions to generate a transformed spectrum; an optimal deformation condition derivation step of deriving optimal deformation conditions from the generated deformation spectrum; The deformation condition is derived by The concentration prediction model is a neural network or a concentration prediction algorithm that is trained to calculate predicted concentration values from an optimally transformed spectrum obtained by transforming the fundamental spectrum under the optimal transformation conditions and a corresponding measured concentration value through machine learning; The optimal deformation condition deriving step includes: A concentration prediction method including calculating the standard deviation of the reference similarity of the transformed spectrum generated in the spectral transformation and combination step, the correlation coefficient with the measured concentration, and the multivariate ratio before and after the spectral transformation, and deriving the transformation conditions that maximize the sum of these as the optimal transformation conditions.

5. The optimal deformation condition deriving step includes: the standard deviation (Ak) of the reference similarity, which is the standard deviation between the reference similarities of the transformed spectra for each transformation condition generated in the spectral transformation and combination step, the correlation coefficient (Bk) with the measured concentration, and the multivariate ratio (Ck) before and after the spectral transformation are calculated, and the transformation condition that maximizes the value of a linear function of these, Fk(Ak, Bk, Ck) = aAk + bBk + cCk (a, b, c are constant values), is derived as the optimal transformation condition; 5. The concentration prediction method according to claim 4, wherein the reference similarity for calculating the standard deviation of the reference similarity, the correlation coefficient with the measured concentration, and the multivariate ratio before and after the spectrum transformation are calculated by the following equations 1, 2, and 3, respectively. [Equation 9] (α is the combined magnitude combination of the transformed spectra, and β is the combined magnitude combination of the reference spectra.) [Equation 10] (X is the reference similarity, Y is the actual concentration of the analyte, Cov(X,Y) is the covariance between the reference similarity (X) and the corresponding concentration of the analyte (Y), and Var(X) and Var(Y) are the variances of the reference similarity (X) and the corresponding concentration of the analyte (Y).) [0011] (The number of multivariates is the number of data points acquired in a given wavelength band corresponding to the X-axis of the acquired spectrum.)

6. 1. A concentration prediction system for predicting a concentration of an analyte, comprising: a spectral data acquisition unit that acquires a basic spectrum of a substance for generating learning data or a predicted spectrum, which is a spectrum of a substance whose concentration is to be measured; a substance concentration data acquisition unit that acquires already acquired substance concentration data and an actual measured concentration of a substance to be measured using a known concentration measuring device; a calculation and control device that generates a modified spectrum of the basic spectrum and the predicted spectrum, calculates optimal modification conditions, performs machine learning, and generates a concentration prediction model; a memory for storing the fundamental spectrum of the analyte obtained by the spectrum data acquisition unit, the measured substance concentration value obtained by the substance concentration data acquisition unit, and the concentration prediction model generated by the calculation and control device; A concentration prediction system comprising: A concentration prediction system in which the optimal transformation conditions are determined by calculating, for the transformed spectra, the standard deviation of the reference similarity, the correlation coefficient with the measured concentration, and the multivariate ratio before and after the spectral transformation, and deriving the transformation conditions that maximize the sum of these as the optimal transformation conditions.

7. The calculation and control device The spectra acquired in the spectrum data acquisition unit are spectrally transformed and combined based on a spectrum transformation algorithm; Among the transformed spectra, the transformed spectrum that shows the best accuracy of concentration prediction is derived as the optimal transformed spectrum; performing machine learning using the optimal deformation spectrum and the corresponding measured concentration of the substance as learning data to generate a concentration prediction model; The concentration prediction system according to claim 6 , wherein the predicted spectrum is input into a generated concentration prediction model to predict the concentration.

8. The memory includes: a spectrum transformation and combination module that transforms the spectrum acquired by the spectrum acquisition unit based on a transformation condition; an optimal spectrum derivation module that acquires an optimal spectrum that indicates the accuracy of optimal concentration prediction from the deformation spectra and derives optimal deformation conditions; a machine learning module that uses the optimal deformation spectrum and the measured substance concentration as learning data to perform machine learning and generate a concentration prediction model; a concentration prediction module that calculates a concentration prediction value using the predicted spectrum as input data; The concentration prediction system according to claim 6 , comprising:

9. a spectral transformation module that transforms a predicted spectrum of a substance whose concentration is to be predicted to generate a transformed spectrum; an optimal spectrum derivation module that calls the spectrum modification module and modifies the predicted spectrum based on optimal modification conditions to calculate an optimal modified predicted spectrum; a density prediction module that receives the optimal deformation predicted spectrum and calculates a predicted density value; comprising The concentration prediction module A computer-generated storage medium storing a concentration prediction algorithm generated by machine learning based on the spectrum transformed based on the optimal transformation conditions and the corresponding measured concentration of a substance, The optimum deformation condition is A computer-readable recording medium that calculates the standard deviation of the reference similarity, the correlation coefficient with the measured concentration, and the multivariate ratio before and after spectral deformation for the deformed spectra, and derives the deformation conditions that maximize the sum of these as the optimal deformation conditions.

10. 1. A concentration measuring device for predicting the concentration of an analyte, comprising: a spectral data acquisition unit for acquiring a predicted spectrum, which is a spectrum of an analyte; a memory device including a spectrum transformation module that transforms a predicted spectrum based on optimal transformation conditions and a concentration prediction module that calculates a predicted concentration value from the optimally transformed predicted spectrum obtained by transforming the predicted spectrum based on the optimal transformation conditions; an arithmetic and control device that reads the concentration prediction module mounted on the memory device and controls the module to calculate a predicted concentration value from a predicted spectrum for which a concentration is to be predicted; A concentration measuring device comprising: The optimum deformation condition is A concentration measuring instrument that calculates the standard deviation of the reference similarity for the predicted spectra, the correlation coefficient with the measured concentration, and the multivariate ratio before and after spectral transformation, and determines the transformation conditions that maximize the sum of these as the optimal transformation conditions.

11. The concentration measuring device is It further comprises a data connection section for connecting an external memory device or receiving input / output of data from an external device, The memory device is The concentration measuring device of claim 10 , which is removably connected to the data connection portion.

12. It further comprises a data connection section for connecting an external memory device or receiving input / output of data from an external device, The memory device is 11. The concentration measuring instrument according to claim 10, wherein a spectrum transformation module that transforms the predicted spectrum based on optimal transformation conditions and a concentration prediction module that calculates a predicted concentration value from the optimally transformed predicted spectrum obtained by transforming the predicted spectrum based on the optimal transformation conditions are received from an external device via the data connection unit and stored.

13. The concentration measuring device according to claim 11 , wherein the data connection unit is integrally configured with the spectral data acquisition unit.

Images